An Uncertainty Aware Semi-Supervised Federated Learning Framework for Radar-based Hand Gesture Recognition

Neural network-based (NN) radar gesture recognition sensors are operated in different domains. The NNs can be trained in a centralized fashion, by using datasets from different individuals and environments. Centralized training faces challenges, such as the risk of data privacy leakage. Federated learning (FL) is a distributed optimization field where data collection, processing, and training of the NN are carried out across multiple clients. A common assumption in FL is that all clients can access ground truth labels. In realistic scenarios, the clients possess partially labeled data, or only a fraction of clients has labeled data. The challenge is known as semi-supervised federated learning (SSFL). For SSFL, one issue is the dependence on the quality of the unlabeled data. In this work, we present a radar-based SSFL framework based on probabilistic pseudo-labeling. It is shown that our framework counteracts poor quality data in the unlabeled dataset during training in gesture sensing.